Fujitsu plans to develop an NPU fabricated on an advanced 1.4nm process by Rapidus, according to a Nikkei Asia report published today. The chip will be designed for AI inference in servers and related systems, with Japan's New Energy and Industrial Technology Development Organization (NEDO) expected to cover approximately two-thirds of the estimated ¥58 billion ($363 million) initial development cost. The project would see the NPU made entirely domestically in terms of both design and manufacturing.
NPUs are dedicated AI inference processors distinct from general-purpose GPUs that dominate AI training. While GPUs excel in the parallel processing required to train LLMs, inference tasks perform better on NPUs, which handle calculations more efficiently. You’ll typically see NPUs in consumer devices like PCs and smartphones, but Fujitsu intends to deploy them in server systems.
